Report CopyRight/DMCA Form For : Detecting Atmospheric Rivers In Large Climate Datasets
Detecting Atmospheric Rivers in Large Climate Datasets Surendra Byna Lawrence Berkeley Laboratory 1 Cyclotron Road 50B 3238 Berkeley CA 94720 USA SByna lbl gov Prabhat Lawrence Berkeley Laboratory 1 Cyclotron Road 50F 1650 Berkeley CA 94720 USA prabhat hpcrd lbl gov Michael F Wehner Lawrence Berkeley Laboratory 1 Cyclotron Road 50F 1650
can occur in as few as five days 1 Their intensity creates We designed an efficient algorithm for identifying AR. a possibility of flooding and wind damage yet at the same using total column integrated precipitable water vapor. time they provide a significant amount of the fresh water data from either observations or simulations. needed for the western states water management systems Our algorithm is highly parallelizable we demonstrate. Although current research is focused on AR events making efficient parallel scaling on a large 1TB dataset. landfall on the western coast of North America the We verify the results from our algorithm against. phenomena is not limited to the northeastern Pacific and published studies by using a set of satellite data that. can occur in other ocean basins have not been previously used for this purpose The. This study of atmospheric rivers is part of on going efforts data used in this study is from Advanced Microwave. to understand the mechanisms responsible for severe but Scanning Radiometer AMSR E satellite described in. infrequent weather events In some winter time events Section 4 We obtain classification accuracy of 92. such as the atmospheric rivers several planetary scale 2 Related Work. conditions must be in phase for such large entrainments of In this section we briefly review related work on. tropical moisture 10 To reach general conclusions many atmospheric rivers and feature detection algorithms. such events must be analyzed individually and as a whole. set To analyze these events they must first be identified. 2 1 Atmospheric Rivers, An atmospheric river is a long and narrow structure in. In this work we develop an efficient algorithm for atmosphere that transports tropical moisture to the far flung. identifying the atmospheric rivers from both observational regions outside of the tropical zone 1 10 Zhu and. data from satellite measurements and climate model output Newell were the first to name this phenomenon. data Part of our motivations is to understand the statistical atmospheric river noting that they typically transport. behavior of the events to analyze how they might change in more water than the Amazon 16 As they can be highly. a warmer climate Hence a key objective is to develop an localized river is an apt description of such a narrow. efficient algorithm to identify AR from large volumes of stream of moisture moving at high speeds across thousands. data allowing us to determine the frequency and intensity of kilometers AR events occur in oceans around the globe. of AR events Additional information about the structure of including the Atlantic basin affecting the British Isles1. AR events notably landfall location intensity and duration. The key characteristic recognized in earlier studies of ARs. are also obtainable by our method and will prove useful in. is the moisture flux 17 However that quantity turns out. projection of future climate change, to be a hard to directly observe In 2004 Ralph et al 11. Observed precipitation and offshore wind speed 8 have established a much simpler set of conditions for identify. been used to identify an AR in the western Pacific basin by atmospheric rivers in satellite observations Their detection. constructing a scatterplot of high quality hourly works with two dimensional data over a uniform mesh on. precipitation and wind data collected at key coastal weather the global and is primarily based on the Integrated Water. stations 14 This ad hoc method is based on setting Vapor IWV content which measures the total water. thresholds of precipitation and wind speed in the upslope content measured in volume in the volume of atmosphere. direction and has proved useful in identifying recent above a unit of earth surface This quantity is measured in. atmospheric river events However this detection method millimeters mm or centimeters cm More specifically. is localized by definition and requires ancillary data such they identify atmospheric rivers as atmospheric features. as total precipitable water from satellite measurements to with IWV 2cm more than 2000 km in length and less. characterize the atmospheric river event Furthermore as than 1000 km in width Based on this definition Ralph and. atmospheric rivers can happen in any ocean basin the colleagues have identified hundreds of atmospheric river. scheme would fail if the event does not make landfall events in the data produced by Special Sensor Microwave. where quality observations are available This likely Imager SSM I satellite observations 1 7. precludes analyses in a global context Application to. In this work we will use a different set of satellite data as. climate simulations may also pose problems due to model. well as output data from a state of the art high resolution. bias in precipitation and wind fields For example the. climate model The observational data we use is from a. thresholds appropriate to observations may not work well if. satellite called Advanced Microwave Scanning Radiometer. model extreme precipitation and winds exhibit systematic. AMSR R This device measures IWV allowing us to use. errors In this paper we present an alternative detection. the same conditions as proposed by Ralph et al 11, scheme based on examining basin wide data characteristic. of the atmospheric river phenomenon Our methodology Objective identification of atmospheric river events is a. allows for detection and characterization of such events in challenging task Identifying observed events in the. both satellite measurements and climate model output As historical record for case study analyses can exploit. such it will prove a critical tool in the projection of future. changes in this class of extreme storm The key,contributions of this work are as follows 1. http cimss ssec wisc edu goes blog archives 3838, associated information such as on shore extreme information to arrive at a final assignment for each. precipitation and wind direction to identify candidate provisional label The most efficient data structure for. events Large scale structural information can then be keeping track of the label equivalence information is called. gained by analyses of satellite measurements 10 union find 2 and the most efficient implementation of the. However analyses of the statistical behavior of union find data structure is an implicit data structure that. atmospheric rivers are also necessary to understand the uses a single array 15 An efficient union find. more general relationship to large scale climatic variations implementation is critical to the overall effectiveness of the. The ability of climate models to simulate atmospheric river two pass algorithm To keep the computational complexity. statistics is key to projecting if these phenomena change as low we chose to keep the binary image in a 2 D array. the climate warms Hence an atmospheric river, identification scheme that neither misidentifies nor misses 3 OUR APPROACH. candidate events is critical to the statistical analysis of Our algorithm processes 2 D meshes defined over the. climate models and their comparison to the observed globe These meshes are relatively small for example the. recent past satellite observation data is defined on a 1 4 mesh with. just over 1M mesh points and the climate model output. 2 2 Feature Detection on Mesh Data, uses a 1 2 mesh Even with fine meshes at 1 10 mesh the. Climate Model and satellite output are typically generated. data associated with a single variable i e integrated water. or regrided on a regular mesh over the globe Following. vapor IWV can easily fit into main memory While we. the methodology used by Ralph et al we perform our. need to process many time steps in the complete dataset. detection on 2 D data on the latitude longitude mesh 11. this can be done in parallel, An atmospheric river is an event that can last for a few. days Our detection algorithm processes one day at a time A schematic illustration of the parallel algorithm for AR. For each day s data the AR appears as a connected region detection is shown in Figure 2 We divided the algorithm. in space where the integrated water vapor content is high into an I O phase and a compute phase The I O phase. This type of feature in space is commonly known as region includes reading the input filenames and vapor data The. of interest Identifying such regions of interest is a basic computation phase consists of thresholding connected. operation in many computer visual analysis tasks component labeling and verification steps Each process. generates an output indicating the presence or absence of. Our detection algorithm proceeds in three steps The first. an AR Our design allows each process to run, step performs a thresholding operation based on IWV. independently without any need for inter process, value mesh points with high IWV values are marked for. synchronization or communication, further processing The second step connects the marked. mesh points into regions This step employs a connected 3 1 I O Phase. component labeling algorithm The connected regions are Our current implementation requires a list of data file. passed to the last step for verification of sizes The first and names to process This list is currently stored in a single. last steps are relatively straightforward In this section we shared file Currently all processes read the file it is. briefly review the algorithms used for connected possible to split this file in the future to reduce metadata. component labeling 3 15 overhead Once each process determines what file to. The IWV data processed by our feature identification process it then proceeds with reading the IWV data. procedure is stored as a 2 dimensional array The output The function that performs the reading of IWV data takes a. from the thresholding step can be treated as a binary image number of optional input parameters such as granularity of. where the foreground pixels are mesh points with large climate data type of data format such as gunzip. IWV values and the background pixels are mesh points compressed format netCDF etc the number of time steps. with small IWV values This allows us to use the connected present in one day s data and regions where AR should be. component labeling algorithms developed from image. processing There are a variety of algorithms for this task. For example there are a number of different parallel. approaches 5 12 some methods using specialized,hardware 4 6 Since the image sizes are relatively. modest in our application we choose to perform connected. component labeling using only a single CPU core, To find the connected component labels we use a two pass. algorithm that gathers the connectivity information among. the foreground pixels and then assign the final labels to. each pixel The two pass algorithms avoid scanning the. image multiple times by manipulating the label equivalence. Figure 2 AR detection tool implemented with MPI, detected This flexibility allows us to detect AR in any operation to set one representative pointing to the other. region of the world at different granularity We choose to have the representative with larger numerical. value pointing to the representative with smaller value The. 3 2 Compute Phase union find data structure can be interpreted as representing. a forest of union find trees where the representative is. 3 2 1 Thresholding the root of each tree Pictorially this is illustrated in Figure. In the compute phase the first step is a set of thresholding 3 By choosing to use non negative integers as labels it is. operations on IWV values Ralph et al 11 specify IWV possible to use the labels as the array index and implement. values 20mm for detecting atmospheric rivers We use the union find data structure in a single array as illustrated. this threshold value for all results reported in the paper in Figure 3. However our detection tool can take on a pair of, thresholds that define the lower bound and an upper bound Using an array to implicitly represent the union find trees. for the IWV values This additional flexibility can be has the advantage that the memory for the union find data. useful for with systematic biases The output of the structure is consecutive in memory Furthermore the find. thresholding step is a collection of mesh points that satisfy operations always traverse to the left in Figure 3 This. the threshold criteria These foreground pixels are then predictable pattern reduces the average cost of the memory. processed by the Connected Component Labeling CCL accesses which improves the overall effectiveness of the. step labeling algorithm, 3 2 2 Connected Component Labeling 3 2 3 Verification. Our connected component labeling implementation is After the connected component labeling step each. based on a two pass algorithm 15 The algorithm can be connected group of mesh points receives a unique label for. broken down into three steps The first step assigns a identification We then compute the length and width of. provisional label to each mesh point visited These each group and impose the relevant constraints i e. provisional labels may turn out to be assigned to connected Length 2000km and Width 1000km 11 in the. mesh points We say that these labels are equivalent This verification step To compute the length we find the. label equivalence information is recorded in a data medial axis of a connected component label Since each. structure called union find The second step works with the pixel on the globe has a relatively constant area by. union find data structure to determine the final label for counting the number of pixels in the connected component. each provision label The third step replaces the provisional label we compute the area and then the average width by. labels with their final values This third step is a series of dividing the area with the medial axis length The detection. straightforward assignments tool classifies an AR event that satisfies all the criteria. The first step examines each mesh point in turn A mesh 4 EXPERIEMENTAL METHODOLOGY. point failing the thresholding conditions will receive a Thus far we have described the algorithm for detecting. special label say 0 to indicate that it is not of interest A atmospheric rivers We are interested in evaluating the. mesh point satisfying the thresholding conditions will performance of our algorithm along the following metrics. receive a provisional label This assignment proceeds as. follows If there is no neighbor with a provisional label How well does our algorithm perform What is its. already then this mesh point receives a new label If any of accuracy. its neighbors have already received a label any of their How well does the implementation scale with large. labels can be assigned to the current mesh point Because data weak scaling. the neighbors are connected to this mesh point and to each How well does the implementation scale with number. other their labels should be the same We say that these of processes strong scaling. labels are equivalent and choose the smallest labels as the We conducted our experiments on the NERSC Cray XE6. representative of the groups of equivalent labels supercomputing system Hopper The system has 6 400. compute nodes with 24 cores total 150 000 cores 2, The union find data structure stores the label equivalence. twelve core AMD MagnyCours 2 1 GHz processors per,information This data structure supports two key. node and 32GB memory per node We used all 24 cores of. operations called union and find Given any provisional. a node for our tests and have one MPI process on each. label the find operation locates its representative Given. core Hopper uses Lustre as its file system with a peak. any two provisional labels the union operation is to record. theoretical I O bandwidth of 35GB s The Lustre system is. that they are equivalent to each other This operation can be. configured with 156 Object Storage Targets OSTs We. implemented as two find operations followed by an, now describe our experimental methodology for addressing. these questions, Figure 3 An array representation of the rooted trees. 4 1 Accuracy of our Approach The strong scaling refers to the ability of an algorithm to. After considering a number of approaches to validate the take advantage of more computing resources to complete. accuracy of our detection algorithm we settled on the same task In our case we keep the input data size fixed. comparing our results to the published AR events in the at 1TB and increasing the number of processes from 100. west coast US by a number of other researchers 1 7 200 500 1 000 2 000 5 000 to 10 000 MPI processes. These papers contain an exhaustive list of atmospheric This data set has 10 000 days of global climate modeling. rivers reaching the US west coast from the year 1998 2008 data therefore we test scaling up to 10 000 processes. We treat the results reported in Dettinger et al 1 from. June 2002 and 2008 as ground truth We note that our 4 4 Data. results are obtained from a different satellite Advanced. Microwave Scanning Radiometer AMSR E satellite 4 4 1 Observational Data. http www ssmi com We use a geophysical dataset derived from observations. collected by the AMSR E satellite The overall dataset. contains sea surface temperature surface wind speed. 4 2 Weak Scaling atmospheric water vapor cloud liquid water and rainfall. The field of climate modeling is undergoing active rate The orbital data of the satellite is mapped to 0 25. research we expect larger and larger simulation datasets to mesh i e each of the data observations is gridded onto a. be produced in the coming years While the dataset sizes 1440 x 720 matrix The daily data collected by AMSR E. are increasing we also have access to large contains gaps because the satellite cannot cover the whole. supercomputing systems to process the data Hence it is globe in a day To obtain complete data for any given day. important that data analysis programs are able to scale up RSS provides time averaged data using a 3 day moving. as more computing resources are provide for large data window. sets To measure this type of scalability we keep the work. given to each process constant but increasing the number In our atmospheric river detection scheme we use the. of processes across 1000 2000 4000 8000 and 10000 vertically integrated water vapor data from files containing. MPI processes while proportionally increasing the 3 day averages of column integrated water vapor The files. problem sizes from 50GB to 1TB We will report both the are compressed into gzipped format gz We converted. time to read the input data and the time to complete the this compressed files into netCDF format The size of each. computations In measuring the I O performance we will 3 day average file in netCDF format is 40 MB In our. report the I O throughput instead of the more common read tests we used observation data for 3100 days which. or write speed There is no synchronization among the amount to 124 GB This dataset is used for verifying the. processes therefore the I O operations on each process are accuracy of our tool in detecting atmospheric rivers in the. not coordinated coastal areas of California Oregon and Washington states. We compare the results with the manually identified list of. 4 3 Strong Scaling AR events by Dettinger et al 1, Figure 4 Some typical atmospheric river events detected by our algorithm from the observational dataset Shown is total column. integrated precipitable water in mm Note that the structure of each event is unique Also note that data irregularities in the. satellite measurements seen as abrupt discontinuities e g in the 2007 12 04 event do not have an adverse effect on the detection. 4 4 2 Model Data, We use climate data generated by the finite volume version. of the Community Atmospheric Model fvCAM in our, scalability study 13 The fvCAM uses a finite volume. approximation to the atmospheric equations of motion and. have been specifically optimized for parallel execution. Output data in netCDF format includes multiple variables. such as pressure humidity temperature total vertically. integrated water vapor For detecting atmospheric rivers. we use the data value for the integrated water vapor The. data is arranged in a 361 x 576 mesh which represents 0 5. Figure 5 Samples of undetected events by our program. latitude by 0 625 longitude There are 4 simulated time. steps per one day i e one per every 6 hours and the total from Dettinger et al s list that did not make landfall The. dataset contains 15 simulated years worth of data that resulting accuracy of our tool is 92 Figure 5 shows a. amounts nearly 450 GB The dataset is stored into 1095 sample of rivers classified by Dettinger et al but not. files and each file consists of 5 days worth of data detected by our algorithm Our tool detected only the. events that reached the western states of the US as we set. To avoid dealing with intra day variations our detection the states as the region of interest These events have vapor. algorithm works with daily averages calculate from the 6 below threshold in some parts of the narrow band and some. hour time steps within the day Since model data does not are wider than 1000km Since these connected labels do. have any missing data we did not need to compute the not fit in the source destination length and width criteria. average for 3 days as in the observational data In our they are not detected as AR by our tool. strong scaling tests we used data related to 10 000 days. which is 1 TB In the weak scaling experiments the data Figure 6 shows statistics of AR events between 2002 and. size is increased in proportion with the number of 2010 For year 2002 the data is available from June to. processes used In these each MPI process analyzes data December and for all other years the events are for the. related to one day For example in a 10 000 process MPI whole year We counted consecutive days with an AR as. job the application processes 10 000 days worth of data one event We separate the AR events in the winter time. which is in the range of 1 TB Similar to the observational from summer months This relative distribution is quite. data analysis in both weak scaling and strong scaling similar to those reported in earlier studies 1 7. studies we analyzed data related coastal areas of, California Oregon and Washington states The tool can be 5 2 Weak Scaling. used for detecting AR in any region by changing the Figure 7 shows results from our weak scaling experiment. longitude and latitude bounding box parameters and the The x axis shows number of MPI processes and the y axis. AR detection criteria shows the time in seconds in logarithmic scale To recall. the experimental setup each process analyzes data for a. 5 RESULTS single day as more processes as added the detection. We outlined three questions to address in our performance algorithm works on a proportionally larger number of days. study In this section we report our findings on each We observe that majority of the execution time of our tool. separately is dominated by I O 98 Since each process only. works on one day s worth of data we expect that the I O. 5 1 Classifier Performance time and the computation time to remain constant as the. We applied our AR detection tool to the observational data. and compare the detected events with the published paper. by Dettinger et al 1 We use the same thresholds listed in. their paper water vapor 20mm length 2000km and,width 1000km and spatial constraints of examining. ARs originating in the tropics and making landfall on the. western US coast Figure 4 shows a sampling of detections. from our program,Our tool detected 81 of the AR events reported in. Dettinger et al Upon further examination we discovered. that Dettinger et al were reporting ARs that were wider. Figure 6 Yearly statistics of atmospheric events from. than 1000km and the rivers that did not actually make. observational data from http www remss com amsre, landfall but was close to it We thereafter removed entries. measured time when 100 processes were used For,Figure 7 Weak scaling times. Figure 8 I O performance with weak scaling, number of processes increases While these costs stayed. relatively constant we noticed a small increase in the instance if the I O overhead with 100 processes is t then. observed time We attribute this to a small fraction of MPI the I O overhead with 200 processes is t 2 and that with. processes taking longer than others to finish their 500 processes is t 5 and so on The combination of reading. processing As the number of processes increase from 1000 file names from input file and the reading vapor data from. to 10000 processes the observed computation time climate data set dominate the overall execution time The. fluctuates between 0 7ms and 1 1ms we believe that these total I O time constitutes 99 of the execution time. fluctuations are random in nature and not systematic. In Figure 9 we see that the computation time speedup. The I O time also increases slightly the main reason for generally agrees with ideal scaling This suggests that the. this increase is due to shared access to the same input file computations are relatively load balanced and amenable to. for reading filenames As the number of processes increase parallelization In this case each process handles data from. the time to read the file names increase from 0 29s for 1000 a number of different days which minimizes the effect of. processes to 1 54s for 10 000 processes In the same tests random fluctuations discussed earlier. the time to read the integrated water vapor data remains. The I O times are very close to ideal speedup for the test. about the same 1 01s 1000 processes and 1 32s for 10 000. cases with 100 200 and 500 processes As indicated, before five processes read from a single data file and their. In Figure 8 we show the aggregate I O throughput against read operations are most likely served by a single disk read. the number of processes We calculated the aggregate I O which means that 100 OSTs can serve 500 processes In. throughput as the sum of I O throughput at each process going from 100 to 500 processes our program is. Since each process runs independently without any effectively using more OSTs from the file system therefore. synchronization measuring global I O bandwidth for the the I O time scales well As more processes are used it is. application does not reflect I O performance of the tool As no longer possible to have each OST serve five processes. the number of processes increases the I O throughput also This creates I O contention and increases the time needed. increases In the model dataset each file contains five days to complete the I O operations We see that the I O time in. of data and is therefore shared by five processes Since Figure 9 goes above the expected value for ideal speedup. each read request fetches a full stripe of data 1 MB in when more than 1000 processes are used. Lustre file system configured on Hopper this 1 MB stripe. In Figure 10 we show the aggregate I O throughput as the. of data includes all water vapor data for all five days This. number of processes increases with fixed data size The I O. explains achieving good I O performance even though each. throughput increases as the number of processes increase. process is expected to read only about 128 KB of data. up to 500 processes following the linear trend From the. 1000 processes case the I O throughput falls short of ideal. 5 3 Strong Scaling, Figure 9 shows strong scaling results with a fixed data size. related to 10 000 days This experiment shows how the. application scales as the number of processes increase. while the data size is fixed The two bars show the I O time. and the computation time The sum of these two costs is. equal to the total execution time of the algorithm The. upper trend line dashed refers to the I O time if ideal. speedup were achieved and the lower trend line represents. computation time if ideal speedup were achieved We. calculated the time with ideal speedup in reference to the. Figure 9 Strong scaling times, Analysis and Processing page 322 Washington DC USA. 1999 IEEE Computer Society, 3 M B Dillencourt H Samet and M Tamminen A general. approach to connected component labeling for arbitrary. image representations J ACM 39 2 253 280 1992, 4 H Flatt S Blume S Hesselbarth T Schunemann and P. Pirsch A parallel hardware architecture for connected. component labeling based on fast label merging In ASAP. pages 144 149 IEEE Computer Society 2008, Figure 10 I O Performance with strong scaling 5 J Greiner A comparison of parallel algorithms for. connected components In SPAA 94 pages 16 25 New, growth Nevertheless the aggregate throughput still York USA 1994. increases reaching 4 6 GB s with 10 000 processes 6 C Y Lin S Y Li and T H Tsai A scalable parallel. hardware architecture for connected component labeling In. Our results indicate that our tool is quite well suited for ICIP pages 3753 3756 IEEE 2010. weak scaling Our tool can analyze more data with a larger. number of processes This will be useful in processing 7 P J Neiman F M Ralph G A Wick Y H Kuo T K. Wee Z Ma G H Taylor and M D Dettinger Diagnosis, output data from century scale climate model integrations. of an intense atmospheric river impacting the pacific. northwest Storm summary and offshore vertical structure. 6 CONCLUSIONS observed with COSMIC satellite retrievals Monthly. Atmospheric rivers are a type of rare weather event capable Weather Review 136 11 4398 4420 2008. of transporting large amounts of water from tropical region 8 P J Neiman A B White F M Ralph D J Gottas and S. to elsewhere They are a important source of fresh water as I Gutman A water vapour flux tool for precipitation. well as a cause of severe flooding and wind damage In this forecasting Proceedings of Institution of Civil Engineers. work we have developed an efficient detection tool for Water Management 162 83 94 2009. automatically detecting ARs We use a combination of 9 R E Newell N E Newell Y Zhu and C Scott. thresholding connected component labeling and Tropospheric rivers A pilot study Geophysical. verification steps to check for the presence of ARs Our Research Letters 19 24 2401 2404 1992. implementation was able to successfully detect 92 of 10 F M Ralph P J Neiman G N Kiladis K Weickmann. ARs that make landfall the results were verified against a and D W Reynolds A multiscale observational case study. manually curated results published by Dettinger et al We of a pacific atmospheric river exhibiting tropical. demonstrated good weak and strong scaling for our extratropical connections and a mesoscale frontal wave. implementation We applied our tool to a large 1TB dataset Monthly Weather Review 139 4 1169 1189 2011. on 10 000 cores and completed the processing in 3 11 F M Ralph P J Neiman and G A Wick Satellite and. seconds This fully automated and highly parallelizable tool CALJET aircraft observations of atmospheric rivers over the. will enable climate scientists to effectively tackle large data eastern north pacific ocean during the winter of 1997 98. challenges from next generation climate simulation output Monthly Weather Review 132 7 1721 1745 2004. 12 K B Wang T L Chia Z Chen and D C Lou Parallel, 7 ACKNOWLEDGMENTS execution of a connected component labeling operation on a. linear array architecture Journal of Information Science. This work was performed under the auspices of the U S. And Engineering 19 353 370 2003,Department of Energy DOE by the Lawrence Berkeley. National Laboratory LBNL under contract DE AC03 13 M F Wehner G Bala P Duffy A A Mirin and R. 76SF00098 LBNL and with support from the Office of Romano Towards direct simulation of future tropical. cyclone statistics in a high resolution global atmospheric. Science BER U S Department of Energy This research. model In Advances in Meteorology page 915303 2010, used resources of the National Energy Research Scientific. Computing Center which is supported by the Office of 14 A B White F M Ralph P J Neiman D J Gottas and S. I Gutman The NOAA coastal atmospheric river, Science of the U S Department of Energy under Contract. observatory In 34th Conference on Radar Meteorology. No DE AC02 05CH11231,15 K Wu E Otoo and K Suzuki Optimizing two pass. connected component labeling algorithms Pattern Analysis. 8 REFERENCES Applications 12 2 117 135 2009,1 M D Dettinger F M Ralph T Das P J Neiman and D. R Cayan Atmospheric rivers floods and the water 16 Y Zhu and R E Newell Atmospheric rivers and bombs. resources of California Water 3 2 445 478 2011 Geophysical Research Letters 21 18 1999 2002 1994. 2 L di Stefano and A Bulgarelli A simple and efficient 17 Y Zhu and R E Newell A proposed algorithm for. connected components labeling algorithm In ICIAP 99 moisture fluxes from atmospheric rivers Monthly Weather. Proceedings of the 10th International Conference on Image Review USA 126 3 725 735 1998.